Audit AI systems for security risks

Assess AI/ML system for prompt injection and jailbreak risks

LLM agents or classifiers needing ATLAS-mapped vulnerability assessment pre-deployment.

• · test prompts (JSON array)
• · target type (llm/classifier/embedding)

• · injection signature matches
• · risk score (0.0-1.0)
• · MITRE ATLAS technique mapping

• · Python: ai_threat_scanner.py
• · MITRE ATLAS framework

Test prompts available, authorization for gray-box/white-box access

Skipping authorization, ignoring indirect RAG injection, conflating with app security

• · Static signature matching
• · MITRE ATLAS IDs in output
• · Access-level gating